> ## Documentation Index
> Fetch the complete documentation index at: https://docs.compute.cx/llms.txt
> Use this file to discover all available pages before exploring further.

# Tinker training API

> Run the official Tinker SDK against a dedicated Compute GPU, then keep the checkpoint.

Tinker on Compute gives the official `tinker` SDK a dedicated training URL backed by one GPU. Your local script controls the training loop. Compute provisions the machine, meters it, and tears it down.

<Warning>
  Tinker is in private alpha. The CLI and operator path are shipped, but public product quotes remain closed until the H100 and MI300X canaries pass. A normal public account currently receives `sku_unavailable` when it tries to create a session.
</Warning>

## Why use it

Use fine-tuning when a capable base model still misses the behavior your application needs. Supervised fine-tuning teaches from prompt and ideal-response pairs. Reinforcement learning is useful when you can score an answer more easily than you can write the ideal one.

The first Compute preset runs LoRA training on `Qwen/Qwen3-4B-Instruct-2507`. It supports cross-entropy for supervised work and importance sampling for RL-shaped loops.

## Run a cookbook recipe

Install the current CLI, sign in, and add prepaid credit first:

```bash theme={null}
curl -fsSL https://compute.cx/install.sh | sh
compute setup
compute credits add 10
```

Approved alpha accounts can run the upstream cookbook without exporting a second credential:

```bash theme={null}
git clone https://github.com/thinking-machines-lab/tinker-cookbook.git
cd tinker-cookbook

compute tinker run --max-spend 25 --idle-timeout 45m -- \
  uv run --with tinker --with datasets \
  python -m tinker_cookbook.recipes.sl_loop \
  model_name=Qwen/Qwen3-4B-Instruct-2507
```

`compute tinker run` does five things:

1. Quotes and provisions a Tinker session.
2. Waits for its public training URL to become healthy.
3. Injects `TINKER_API_KEY`, `TINKER_BASE_URL`, and `TINKER_MODEL` into the child process.
4. Runs the command and forwards its exit status and terminal signals.
5. Stops the CLI-owned session and waits for checkpoint harvest.

The SDK key is not printed or added to your outer shell.

## Run your own script

Your script can build the official client from the injected environment:

```python theme={null}
import os

import tinker

service = tinker.ServiceClient(
    base_url=os.environ["TINKER_BASE_URL"],
    api_key=os.environ["TINKER_API_KEY"],
)
```

Then run it through the same lifecycle:

```bash theme={null}
compute tinker run --max-spend 25 -- python train.py
```

`tinker.ServiceClient()` with no arguments also reads the two official environment variables. Do not hard-code the hosted Tinker URL; each Compute session has its own `https://tinker-<id>.api.compute.cx` base URL.

## What you get

The training script receives the normal Tinker futures and results. When the session stops, Compute harvests saved weights to object storage. List the available files with:

```bash theme={null}
compute tinker checkpoints tns_...
```

Download a harvested checkpoint from the session page in the [dashboard](https://compute.cx/tinker). The CLI currently lists checkpoints but does not download them.

<CardGroup cols={2}>
  <Card title="CLI lifecycle" href="/tinker/cli" icon="terminal">
    Choose between one-shot runs, interactive shells, and warm user-owned sessions.
  </Card>

  <Card title="SDK compatibility" href="/tinker/api" icon="code">
    Check the supported model and API boundaries, including the SDK environment and losses.
  </Card>
</CardGroup>